-MathModel started as a bet: could five independent LLM agents coordinate through a shared JSON workspace to solve math modeling problems? The answer was yes, but not in the way I expected.
The architecture. One orchestrator agent reads the problem description, decomposes it into subtasks, and dispatches them to five specialists: DataAnalyzer, ModelBuilder, Solver, Validator, and Reporter. Each specialist reads from and writes to a shared JSON context. The orchestrator monitors the context and decides when to move on.
What worked. The shared JSON context was surprisingly effective. Specialists didn't need to parse each other's natural language — they read structured data. The ModelBuilder wrote equations as JSON arrays; the Solver consumed them directly. This clean interface between agents eliminated most of the hallucination cascades that plague multi-agent systems.
What broke. The orchestrator kept making up tasks that didn't exist. Given a vague problem description, it would invent subproblems that sounded plausible but weren't asked. Mitigation: a strict task-schema validation step before dispatching.
Token budget. Five agents reading and writing to a shared context adds up fast. A typical problem consumed ~50K tokens across all agents. For MCM/ICM competition problems, this was acceptable. For real-time applications, it's not.
The insight. Multi-agent systems don't fail because agents are dumb. They fail because agents don't share enough context — or share too much. The shared JSON workspace hit a sweet spot: structured enough for machines, readable enough for humans debugging why the Solver produced negative probabilities.